Fast and Accurate Radiative Transfer Calculations Using Principal Component Analysis for (Exo-)Planetary Retrieval Models

Abstract:

Radiative transfer (RT) computations form the engine of atmospheric retrieval codes. However, full treatment of RT processes is computationally expensive, prompting usage of two-stream approximations in current exoplanetary atmospheric retrieval codes [Line et al., 2013]. Natraj et al. [2005, 2010] and Spurr and Natraj [2013] demonstrated the ability of a technique using principal component analysis (PCA) to speed up RT computations. In the PCA method for RT performance enhancement, empirical orthogonal functions are developed for binned sets of inherent optical properties that possess some redundancy; costly multiple-scattering RT calculations are only done for those few optical states corresponding to the most important principal components, and correction factors are applied to approximate radiation fields. Kopparla et al. [2015, in preparation] extended the PCA method to a broadband spectral region from the ultraviolet to the shortwave infrared (0.3-3 micron), accounting for major gas absorptions in this region.

Here, we apply the PCA method to a some typical (exo-)planetary retrieval problems. Comparisons between the new model, called Universal Principal Component Analysis Radiative Transfer (UPCART) model, two-stream models and line-by-line RT models are performed, for spectral radiances, spectral fluxes and broadband fluxes. Each of these are calculated at the top of the atmosphere for several scenarios with varying aerosol types, extinction and scattering optical depth profiles, and stellar and viewing geometries. We demonstrate that very accurate radiance and flux estimates can be obtained, with better than 1% accuracy in all spectral regions and better than 0.1% in most cases, as compared to a numerically exact line-by-line RT model. The accuracy is enhanced when the results are convolved to typical instrument resolutions. The operational speed and accuracy of UPCART can be further improved by optimizing binning schemes and parallelizing the codes, work on which is under way.